US 7356504 B2 Abstract A preferred embodiment comprises a method for determining value-at-risk based on tick-by-tick financial data. Major steps of the method comprise the following: (1) financial market transaction data is electronically received by a computer; (2) the received financial market transaction data is electronically; (3) a time series z is constructed that models the received financial market transaction data; (4) an exponential moving average operator is constructed; (5) an operator is constructed that is based on the exponential moving average operator; (6) a causal operator Ω[z] is constructed that is based on the iterated exponential moving average operator; (7) values of predictive factors are calculated; (8) the values calculated by the computer are stored in a computer readable medium, and (9) value-at-risk is calculated from the values stored in step (8).
Claims(11) 1. A method of determining value-at-risk, comprising the steps of:
electronically receiving financial market transaction data over an electronic network;
electronically storing in a computer-readable medium said received financial market transaction data;
constructing an inhomogeneous time series z that represents said received financial market transaction data;
constructing an exponential moving average operator;
constructing an iterated exponential moving average operator based on said exponential moving average operator;
constructing a time-translation-invariant, causal operator Ω[z] that is a convolution operator with kernel ω and that is based on said iterated exponential moving average operator;
electronically calculating values of one or more predictive factors relating to said time series z, wherein said one or more predictive factors are defined in terms of said operator Ω[z];
electronically storing in a computer readable medium said calculated values of one or more predictive factors; and
electronically calculating value-at-risk from said calculated values.
2. The method of
3. The method of
ν depends on a chosen interpolation procedure.
4. The method of
Δ[τ]=γ(EMA [ατ, 1]+EMA[ατ2]−2 EMA [αβτ, 4]), where γ is fixed so that the integral of the kernel of the differential operator from the origin to the first zero is 1; α is fixed by a normalization condition that requires Δ[τ; c]=0 for a constant c; and β is chosen in order to get a short tail for the kernel of the differential operator Δ[τ].
5. The method of
6. The method of
7. The method of
8. The method of
Volatility[τ,τ′, p;z]=MNorm
where
MNorm[τ,p;z]=MA[τ;|z|^{p}]^{1/p}, and
and where p satisfies 0<p≦2, and τ′ is a time horizon of a return r[τ]=Δ[τ; x], where x represents a logarithmic price.
9. The method of
corresponding to a linear interpolation procedure.
10. The method of
ν=1, corresponding to a previous point interpolation procedure.
11. The method of
ν=μ, corresponding to a next point interpolation procedure.
Description This application claims priority to U.S. Provisional Application No. 60/200,742, filed May 1, 2000; U.S. Provisional Application No. 60/200,743, filed May 1, 2000; U.S. Provisional Application No. 60/200,744, filed May 1, 2000; and U.S. Provisional Application No. 60/274,174, filed Mar. 8, 2001. The contents of the above applications are incorporated herein in their entirety by reference. For banks and other financial institutions, risk measurement plays a central role. Risk levels must conform to the capital adequacy rule. An error in the computed risk level may thus affect a bank's investment strategy. The state of the art is measuring risk by analyzing daily data: using one market price per working day and per financial instrument. In this description, the stochastic error of such a risk measure is demonstrated in a new way, concluding that using only daily data is insufficient. The challenge for statisticians is to analyze the limitations of risk measures based on daily data and to develop better methods based on high-frequency data. This description meets this challenge by introducing the time series operator method, applying it to risk measurement and showing its superiority when compared to a traditional method based on daily data. Intra-day, high frequency data is available from many financial markets nowadays. Many time series can be obtained at tick-by-tick frequency, including every quote or transaction price of the market. These time series are inhomogeneous because market ticks arrive at random times. Irregularly spaced series are called inhomogeneous, regularly spaced series are homogeneous. An example of a homogeneous time series is a series of daily data, where the data points are separated by one day (on a business time scale which omits the weekends and holidays). Inhomogeneous time series by themselves are conceptually simple; the difficulty lies in efficiently extracting and computing information from them. In most standard books on time series analysis, the field of time series is restricted to homogeneous time series already in the introduction (see, e.g., Granger C. W. J. and Newbold P., 1977, Forecasting economic time series, Academic Press, London; Priestley M. B., 1989, Non-linear and non-stationary time series analysis, Academic Press, London; Hamilton J. D., 1994, Time Series Analysis, Princeton University Press, Princeton, N.J.) (hereinafter, respectively, Granger and Newbold, 1977; Priestley, 1989; Hamilton, 1994). This restriction induces numerous simplifications, both conceptually and computationally, and was almost inevitable before cheap computers and high-frequency time series were available. U.S. Provisional Application No. 60/200,743, filed May 1, 2000, discloses a new time series operator technique, together with a computationally efficient toolbox, to directly analyze and model inhomogeneous as well as homogeneous time series. This method has many applications, among them volatility or Value-at-Risk (VaR) computations tick by tick. A comparison is made herein between VaR results based on daily data, sampled at a certain daytime, and results based on tick-by-tick data and the new time series operator technique. If using daily data, a surprising and (for practitioners) alarming sensitivity against the choice of the sampling daytime is observed. The stochastic noise seems higher than acceptable to risk managers. An alternative VaR computation based on tick-by-tick data and a new time series operator technique is shown to have similar properties, except for two advantages: distinctly reduced noise and availability of up-to-date results at each tick. The time series operators can also be used in the formulation of old and new generating processes of time series. This opens new ways to develop process equations with new properties, also for inhomogeneous time series. A preferred embodiment comprises a method for determining value-at-risk based on tick-by-tick financial data. Major steps of the method comprise the following: (1) financial market transaction data is electronically received by a computer; (2) the received financial market transaction data is electronically; (3) a time series z is constructed that models the received financial market transaction data; (4) an exponential moving average operator is constructed; (5) an operator is constructed that is based on the exponential moving average operator; (6) a causal operator Ω[z] is constructed that is based on the iterated exponential moving average operator; (7) values of predictive factors are calculated; (8) the values calculated by the computer are stored in a computer readable medium, and (9) value-at-risk is calculated from the values stored in step (8). 1. The Time Series Operator Technique In this description, only a minimum of a description of time series operators is given, so the applications of the following sections can be understood. The theory of the time series operators is explained in U.S. Provisional Application No. 60/200,743 1.1 Inhomogeneous Time Series A time series z consists of elements or ticks z_{i }at times t_{l}. The sequence of these time points is required to be growing, t_{l}>t_{l−1}. A general time series is inhomogeneous, meaning that the sampling times t_{l }are irregular. For a homogeneous time series, the sampling times are regularly spaced, t_{l}−t_{l−1}=δt. For some discussions and derivations, a continuous-time version of z has to be assumed: z(t). However, the operator methods that are eventually applied only need the discrete time series (t_{l}, z_{i}). The letter x is used to represent the time series of logarithmic middle prices, x=(In p_{bid}+1n P_{ask})/2. This quantity is used in the applications. 1.2 Operators An operator Ω, mapping from the space of time series into itself, is depicted in Linear and translation-invariant operators are equivalent to a convolution with a kernel ω:
A causal kernel has ω(t)=0 for all t<0. No information from the “future” is used. If ω(t) is non-negative, ω[z] is a weighted moving average of z whose kernel should be normalized: The range of an operator is the first moment of its kernel: Operators are useful for several reasons, as will be shown. One important aspect is to replace individual ticks from the market by local short-term averages of ticks. This mirrors the view of traders who consider not only the most recent tick but also the prices offered by other market makers within a short time interval. 1.3 The Simple EMA Operator The Exponential Moving Average (EMA) operator is a simple example of an operator. It is written EMA [τ; z] and has an exponentially decaying kernel (as shown in
According to eqs. (3) and (4), the range of the operator EMA[τ; z] and its kernel is
The values of EMA[τ; z](t) can be computed by the convolution of eq. (1), if z(t) is known in continuous time. This implies an integration whose numerical computation for many time points t is costly. Fortunately, there is an iteration formula that makes this computation much more efficient and, at the same time, solves the problem of discrete data. This means that we do not need to know the whole function z(t); we just need the discrete time series values z_{1}=z(t_{i}) at irregularly spaced time points t_{l}. The EMAs are calculated by the following iteration formula:
For a homogeneous time series, μ and ν are constants. A homogeneous time series can alternatively be regarded as a truly discrete time series to which interpolation does not apply. This is mentioned here because it is a popular approach used by traders. For such a discrete time series, t_{n}−t_{n−1 }is defined to be 1, and the following definition is appropriate:
The range of an operator for a genuine discrete time series has a new definition: The iteration equation (6) is computationally efficient, extremely so when compared to a numerical convolution based on eq. (1). No other operator can be computed as efficiently as the simple EMA operator. However, there are means to use the iteration equation (6) as a tool to efficiently compute operators with other kernels, as shown below. An iteration formula is not enough. We have to initialize EMA [τ; z] (t_{0}) at the start of the time series. For this, we can take z_{0}=z (t_{0}) or another typical value of z. This choice introduces an initial error of EMA [τ; z] (t_{0}) which decreases exponentially with time. Therefore, we also need a build-up period for EMA [τ; z]: a time period over which the values of EMA [τ; z] should not yet be applied because of their initial error. Build-up periods should be multiples of τ, e.g., 5τ. The choice of a large enough build-up period is discussed in U.S. Provisional Application No. 60/200,743. 1.4 The Operator EMA [τ, n; z] Time series operators can be convoluted: a time series resulting from a an operator can be mapped by another operator. This is a powerful method to generate new operators with different kernels. The EMA [τ, n; z] operator results from the repeated application of the same simple EMA operator. The following recursive definition applies:
The operator EMA [τ, n; z] has the following kernel: The family of functions of eq. 12 is related to Laguerre polynomials which are orthogonal with respect to the measure e^{−t }(for τ=1). Operators, i.e., their kernels, can be linearly combined. This is a powerful method to generate more operators. Linear combinations of EMA[τ, n; z] operators with different n but identical τ values have kernels that correspond to expansions in Laguerre polynomials. This means that any kernel can be expressed as such a linear combination. The convergence, however, of the Laguerre expansion may be slow. In practice, a small set of useful operators can be prepared with all the kernels needed. Aside from the discussed expansion, it is also possible to linearly combine kernels with different τ values. Some useful types of combined operators are presented in U.S. Provisional Application No. 60/200,743. 1.5 The Operator MA[τ, n; z] The moving average (MA) operator has kernels with useful properties as shown in The formula for the MA[τ, n] kernel is
Many other kernel forms can be constructed through different linear combinations of EMA[τ, n; z] and other operators. 1.6 From Returns to Differentials Most statistics in finance is based on returns: price changes rather than prices. Simple returns have a rather noisy behavior over time; we often want differences between local averages of x: smoothed returns. Smoothed returns are computed by differential operators. Examples:
The expectation value of squared smoothed returns may differ from that of the corresponding simple returns. This has to be accounted for when comparing the two concepts, for example in terms of a factor c in eq. (20). 1.7 Volatility Measured by Operators Volatility is a central term in risk measurement and finance in general, but there is no unique, universally-accepted definition. There are volatilities derived from option market prices and volatilities computed from diverse model assumptions. In this description, the focus is on historical volatility: a volatility computed from recent data of the underlying instrument with a minimum of parameters and model assumptions. For computing the time series of such a volatility, a time series operator is again the suitable tool. We first define the nonlinear moving norm operator:
The volatility of a time series x can now be computed with the help of the moving norm operator: Eq. (17) is based on a moving average (MA) and a differential (Δ) operator. In principle, we may choose any MA and Δ operator according to our preference. In the applications of section 3, this choice is made explicit. The volatility definition of eq. (17), as any definition of historical volatility, necessarily has two timing parameters:
Computing recent volatility is a central ingredient of risk assessment in risk management. Here it serves as an example to demonstrate the usefulness and superiority of time series operators. The RiskMetrics™ method (see J. P. Morgan, 1996, RiskMetrics—technical document, Technical report, J. P. Morgan and International marketing—Reuters Ltd.) is chosen as a well-known example. First it is shown to be a special application of the time series operator technique. Then a better volatility computation method, also based on the time series operator technique, is proposed. Thus two approaches are compared:
In both cases, squared volatility is defined as the expectation σ^{2 }of squared, working-daily changes of the logarithmic middle price x. This is for the sake of a meaningful comparison; it does not imply that using squared returns is necessarily the best choice for an optimal volatility definition. 2.1 Conventional Computation Based on Daily Data The RiskMetrics method is based on an IGARCH model. Its volatility formula gives the conditional expectation of the squared return assuming IGARCH:
Thanks to the regularity of the underlying homogeneous time series, μ=0.94 is a constant. In general, the constancy of μ makes the operator technique particularly efficient for homogeneous time series. In The differences between the two curves are surprisingly large: up to 25%, an alarming uncertainty for risk managers. Risk levels are linked to a bank's capital through the capital adequacy rule, so differences in risk measurements have a major impact on banking. In our case, two risk managers measure very different volatility and thus risk levels for the same financial instrument, just because they live in different time zones. A difference can persist over weeks, as shown in Both deviating volatility values cannot be right at the same time; there must be an error in these values. This error is of a stochastic nature; there is no systematic bias dependent on the daytime. In
The focus here is not so much the behavior of RiskMetrics (IGARCH), but the problems of using homogeneous, daily data in general, no matter which GARCH-type or other model is investigated. The significance of most results can be improved by using all the available information, tick by tick, as shown in the next section. 2.2 Tick-by-Tick Volatility Computation For the sake of a fair comparison, a tick-by-tick volatility computation is introduced that follows RiskMetrics as closely as possible. There are two innovative modifications:
The new volatility measure is again defined in operator notation (where “wdays” stands for working days):
The constant c compensates for the fact that we use smoothed returns x−EMA[τ, 4; x] as introduced in section 2.6 instead of the simple returns of section 3.1. In the case of x following a Gaussian random walk, the theoretically correct value is c=128/93. Using this factor eliminates a systematic bias of the tick-by-tick volatility as compared to the RiskMetrics volatility. Eq. (20) is computed on a special business time scale defined as follows. The 49 weekend hours from Friday 8 pm GMT to Sunday 9 pm GMT are compressed to the equivalent of only 1 hour outside the weekend. This fully corresponds to the time scale of RiskMetrics which omits the weekend days. A more sophisticated and appropriate choice of the business time scale would be the υ-time of Dacorogna et al. (1992) (Dacorogna M. M. Müller U. A., Nagler R. J., Olsen R. B., and Pictet O. V., 1993, A geographical model for the daily and weekly seasonal volatility in the FX market, Journal of International Money and Finance, 12(4), 413-438), but this is avoided here in order to keep the approach as close to RiskMetrics as possible. The variations of the volatility level over time are moderate in The curves of The lower noise level of the tick-by-tick volatility is now plausible, but we need scientific evidence for this. In the general case, such evidence can be gained through Monte-Carlo studies based on a certain process assumption, comparing the error variances of the RiskMetrics volatility and the tick-by-tick volatility. In the case of a Gaussian random walk, we have even stronger evidence: by using continuously-overlapping returns instead of non-overlapping returns, the error variance of the empirically determined σ^{2 }is reduced to ⅔ of the original value. The tick-by-tick operator is indeed using (almost) continuously overlapping returns. In addition to this, it is based on smoothed rather than simple returns, which also leads to a reduction of stochastic noise. Other advantages of tick-by-tick, operator-based methods are the efficient computation based on iterations and the updating at every new tick. Thanks to fast updates, the volatility measure can quickly react to new market events such as shocks, at any time of day. 2.3 VaR Computation in Real-Time Conventional Value-at-Risk (VaR) computations are done once a day, usually in the evening. Traders and portfolio managers typically do not know their current risk; they just know yesterday evening's risk. What they really need is a real-time VaR computation, updated as quickly as they can change their positions. The tick-by-tick operator proposed in section 3.2 and eq. (20) is a tool to make a real-time VaR possible. A real-time VaR computed according to these guidelines would still be somewhat similar to RiskMetrics, except for the substantial benefits of lower noise and a higher updating frequency. There are many criticisms of RiskMetrics that would apply to it, too. Some researchers, for example, replace the IGARCH model by another GARCH-type model. Other researchers focus on the behavior of extreme price changes which may follow other laws than average-size changes. Moreover, return observations over intervals other than daily (for example, hourly or weekly) returns contain valuable information that should also be used in a VaR computation. 3. Process Equations Based on Operators Processes of the ARMA and GARCH families can be expressed in terms of time series operators, as we have seen for IGARCH in eq. (19). The squared conditional volatility of a GARCH(1,1) process, for example, can be written as follows:
This rephrased form of GARCH(1,1) is a starting point of interesting new developments. Initially, it applies to a discrete, homogeneous time series in the sense of eq. (10), but it allows for a direct and efficient computation of the GARCH(1, 1) volatility from inhomogeneous data, since the operator technique is also suited to inhomogeneous time series. Moreover, eq. (22) can be modified to obtain other processes. The kernel of the EMA operator can, for example, be replaced by other kernels. The return x(t′)−x(t′−Δt) can be replaced by a smoothed return computed by a differential operator which reflects the perception of market participants better that the simple return. Dacorogna et al. (1998) (Dacorogna M. M., Müller U. A., Olsen R. B., and Pictet O. V., 1998, Modelling short-term volatility with GARCH and HARCH models, published in “Nonlinear Modelling of High Frequency Financial Time Series” edited by Christian Dunis and Bin Zhou, John Wiley, Chichester, 161-176) have introduced the EMA-HARCH process to model some empirical facts of high-frequency data in finance: the long memory of volatility, the fat tails of the return distribution and the asymmetric causality between fine-grained (high-resolution) and coarse-grained (low-resolution) volatilities as found by Müller et al. (1997) (Müller U. A., Dacorogna M. M., Davé R. D., Olsen R. B., Pictet O. V., and von Weizsäcker J. E., 1997, Volatilities of different time resolutions—analyzing the dynamics of market components, Journal of Empirical Finance, 4(2-3), 213-239). This is one of the first processes whose equation is written with the help of a time series operator:
The “partial volatilities” σ_{j} ^{2 }correspond to market segments and are written in terms of the EMA operator: Most financial markets produce inhomogeneous data, irregularly spaced in time. The time series operators described herein are able to directly use inhomogeneous data to estimate statistical variables such as volatilities. This computation is made efficient by using iteration formulas. The operator technique efficiently works also for homogeneous, equally spaced data. Starting from the simple exponential moving average (EMA) operator, large families of operators with different kernels and different purposes can be constructed. A wider overview of these more complex operators, which are still computationally efficient, is described in U.S. Provisional Application No. 60/200,743. One example is a tick-by-tick Fourier analysis on a moving sample. Thanks to averaging, the operator technique often produces results with less noise (lower stochastic errors) than conventional methods based on homogeneous time series. This is also the case for the main application of the above description: volatility of daily returns as needed for Value-at-Risk (VaR) computations. The conventional RiskMetrics methods have a rather high stochastic error which is demonstrated in a new way: volatility is computed twice with daily data. In one case, the data is always sampled at 7 am GMT (late afternoon in East Asia), in the other case at 5 pm (late afternoon in London). The results of the two computations can differ by some 25% for many days in a row—an embarrassing fact for risk managers. The tick-by-tick alternative of a preferred embodiment based on time series operators does not have this sensitivity against the choice of the sampling time of day and has less noise, while keeping the essential preferred characteristics of the known RiskMetrics method: it is still based on daily returns and still has a range (center of gravity of the kernel) of around 16 working days. The same technique is preferably used to determine a real-time VaR, updated with every new incoming tick from a market, with less noise than the corresponding results from conventional methods. Many methods of calculating VaR from volatility are known in the art (see, for example, Chapter 14: Value at Risk of Options, Futures, and Other Derivatives, by John C. Hull (4^{th }ed. 2000). Finally, the operator technique can be used to formulate time series generation process equations. This is possible for well-known processes such as GARCH and new, more complex processes such as EMA-HARCH. The formulation in terms of operators has many advantages: flexibility and applicability to irregularly-spaced time series. Although the subject invention has been described with reference to preferred embodiments, numerous modifications and variations can be made that will still be within the scope of the invention. No limitation with respect to the specific embodiments disclosed herein is intended or should be inferred. Patent Citations
Non-Patent Citations
Referenced by
Classifications
Legal Events
Rotate |